Machine learning-based prediction of emergency neurosurgery within 24 h after moderate to severe traumatic brain injury

Abstract

Background: Rapid referral of traumatic brain injury (TBI) patients requiring emergency neurosurgery to a specialized trauma center can significantly reduce morbidity and mortality. Currently, no model has been reported to predict the need for acute neurosurgery in severe to moderate TBI patients. This study aims to evaluate the performance of Machine Learning-based models to establish to predict the need for neurosurgery procedure within 24 h after moderate to severe TBI.

Methods: Retrospective multicenter cohort study using data from a national trauma registry (Traumabase®) from November 2011 to December 2020. Inclusion criteria correspond to patients over 18 years old with moderate or severe TBI (Glasgow coma score ≤ 12) during prehospital assessment. Patients who died within the first 24 h after hospital admission and secondary transfers were excluded. The population was divided into a train set (80% of patients) and a test set (20% of patients). Several approaches were used to define the best prognostic model (linear nearest neighbor or ensemble model). The Shapley Value was used to identify the most relevant pre-hospital variables for prediction.

Results: 2159 patients were included in the study. 914 patients (42%) required neurosurgical intervention within 24 h. The population was predominantly male (77%), young (median age 35 years [IQR 24-52]) with severe head injury (median GCS 6 [3-9]). Based on the evaluation of the predictive model on the test set, the logistic regression model had an AUC of 0.76. The best predictive model was obtained with the CatBoost technique (AUC 0.81). According to the Shapley values method, the most predictive variables in the CatBoost were a low initial Glasgow coma score, the regression of pupillary abnormality after osmotherapy, a high blood pressure and a low heart rate.

Conclusion: Machine learning-based models could predict the need for emergency neurosurgery within 24 h after moderate and severe head injury. Potential clinical benefits of such models as a decision-making tool deserve further assessment. The performance in real-life setting and the impact on clinical decision-making of the model requires workflow integration and prospective assessment.

Keywords: Artificial intelligence; Emergency neurosurgery; Prediction models; Trauma; Traumatic brain injury.

Fig. 1

Distribution of train-valid-test dataset: “Synthetic” examples are created by Synthetic Minority Oversampling Technique (SMOTE) only in the training set to cope with the bias toward the majority class due to the imbalanced distribution of the target variable (Emergency neurosurgery within the 24 h hours after admission)

Fig. 2

Flowchart

Fig. 3

Area under the curve of the different artificial intelligence models after the "Test set" phase

Fig. 4

Matrice of confusion of the models: The confusion matrix describes the performance of each classification model. For example, the Catboost model has a balanced prediction with 44 false negatives (patient that will require neurosurgery but not identified by the model) and 68 false positives (patient that will not require neurosurgery but identified by the model as requiring emergency neurosurgery)

Fig. 5

Shapley values of the Catboost model: each point on the summary plot is a Shapley value for a variable and an instance. The color represents the value of the variable from low (blue) to high (red). The most important variable that helps in the prediction is the systolic pressure upon arrival of the physician-staffed EMS and the less important is. For example, the distribution of the feature value along the x-axis indicates that low systolic pressure contributes to a prediction of negative outcome and high systolic pressure contributes to a prediction of positive outcome